John DiNardo 105- What is a good estimate of next quarter’s GDP?
- Does this structural model of the US labor market provide representation
adequate enough for the purposes of evaluating potential policies? - What are the causes and consequences of black culture?
In my experience, what type of questions one is interested in asking often suggests
what type of statistics one finds useful. While both types of questions are routinely
asked by economists, the types of problems entailed seem very different to me
(even if they do not appear this way to some Bayesians). This is not to imply that
the second set of questions are necessarily illegitimate: I wouldn’t want to suggest
that people stop trying to estimate next quarter’s GDP!
Indeed, when and where probability and statistics are most “useful” is one sub-
ject which divides many Bayesian and non-Bayesians and one that we explore in
section 3.3.2.
3.2 Six surprising ideas and one puzzle
It may seem hard to believe that one’s views on the metaphysics of statistics have
consequences. In this section I enumerate six “surprising ideas” that I think go
to the heart of many differences between non-Bayesians and Bayesians. For my
purposes, I will focus on suggestions for practice that are most frequently invoked
by Bayesians or radical subjectivists that are at furthest remove frommy ownnon-
Bayesian views. Despite this, my goal isn’t to criticize them. Indeed, if they strike
youas sensible, perhaps you are a (closet) Bayesian!
3.2.1 Six surprising ideas
- The absence or presence of data-mining strategies, specification mining, non-
random sampling, or non-random assignment are (should be) irrelevant to the
inference of a set of data. Put differently, what could have happened, but didn’t,
in an experiment should make no difference to the evidential import of the
experiment:
considerations about samples that havenotbeen observed, are simply not
relevant to the problem of how we should reason from the one that has been
observed. (Jaynes, 1976, p. 200)
Unbiased estimates, minimum variance properties, sampling distributions,
significance levels, power, all depend on something...that is irrelevant in
Bayesian inference – sample space. (Lindley, 1971, p. 426)- Pre-specified research design is a waste of time:
In general, suppose that you collect data of any kind whatsoever – not neces-
sarily Bernoullian, nor identically distributed, nor independent of each other- stopping only when the data thus far collected satisfy some criterion of a